Sustained Resilience via Live Process Cloning

More flexible fault tolerance approaches with lower overhead are a must for the next generation of supercomputers that rely on massive numbers of computational elements. This work proposes a reactive method for fault resilience in high-performance computing (HPC) systems based on forward execution instead of rollback to checkpoints. We study the feasibility of combining redundancy with live process cloning to create highly reliable HPC systems. The main motivation is to avoid costly checkpoint restart approaches. We present live process cloning as a mechanism to create a copy of a running process on-the-fly. We show that the reliability of a dual redundant system with live process cloning is as good as a triple redundant system even for very large systems. We also investigate the effect of node failure and the changes in Mean time to Interrupt (MTTI) of the application. This provides a better understanding of the available time to recover from a failure by cloning a healthy replica.